Rename arguments `--disable-nan-detection` to `--enable-nan-detection` (#2066)

python/sglang/srt/layers/sampler.py

@@ -29,7 +29,7 @@ logger = logging.getLogger(__name__)
 class Sampler(nn.Module):
     def __init__(self):
         super().__init__()
-        self.use_nan_detectioin = not global_server_args_dict["disable_nan_detection"]
+        self.use_nan_detectioin = global_server_args_dict["enable_nan_detection"]
 
     def forward(
         self,

python/sglang/srt/managers/schedule_batch.py

@@ -57,7 +57,7 @@ global_server_args_dict = {
     "triton_attention_reduce_in_fp32": ServerArgs.triton_attention_reduce_in_fp32,
     "disable_mla": ServerArgs.disable_mla,
     "torchao_config": ServerArgs.torchao_config,
-    "disable_nan_detection": ServerArgs.disable_nan_detection,
+    "enable_nan_detection": ServerArgs.enable_nan_detection,
     "enable_dp_attention": ServerArgs.enable_dp_attention,
 }
 

python/sglang/srt/model_executor/model_runner.py

@@ -139,7 +139,7 @@ class ModelRunner:
                 "disable_mla": server_args.disable_mla,
                 "torchao_config": server_args.torchao_config,
                 "disable_penalizer": server_args.disable_penalizer,
-                "disable_nan_detection": server_args.disable_nan_detection,
+                "enable_nan_detection": server_args.enable_nan_detection,
                 "enable_dp_attention": server_args.enable_dp_attention,
             }
         )
@@ -276,6 +276,10 @@ class ModelRunner:
             else None
         )
         self.dtype = self.vllm_model_config.dtype
+        if self.sliding_window_size:
+            assert (
+                self.server_args.attention_backend == "flashinfer"
+            ), "Only flashinfer supports window attention."
 
         logger.info(
             f"Load weight end. "

python/sglang/srt/models/gemma2.py

@@ -332,6 +332,7 @@ class Gemma2ForCausalLM(nn.Module):
     # Gemma does not apply LoRA to the embedding layer.
     embedding_modules = {}
     embedding_padding_modules = []
+    supports_lora = True
 
     def __init__(
         self,

python/sglang/srt/server_args.py

@@ -124,7 +124,6 @@ class ServerArgs:
     disable_custom_all_reduce: bool = False
     disable_mla: bool = False
     disable_penalizer: bool = False
-    disable_nan_detection: bool = False
     enable_overlap_schedule: bool = False
     enable_mixed_chunk: bool = False
     enable_dp_attention: bool = False
@@ -132,6 +131,7 @@ class ServerArgs:
     torch_compile_max_bs: int = 32
     cuda_graph_max_bs: int = 160
     torchao_config: str = ""
+    enable_nan_detection: bool = False
     enable_p2p_check: bool = False
     triton_attention_reduce_in_fp32: bool = False
     num_continuous_decode_steps: int = 1
@@ -171,11 +171,11 @@ class ServerArgs:
         else:
             gpu_mem = get_nvgpu_memory_capacity()
         if gpu_mem < 25000:
+            self.chunked_prefill_size //= 4  # make it 2048
+            self.cuda_graph_max_bs = 4
             logger.warning(
                 "Automatically adjust --chunked-prefill-size for small GPUs."
             )
-            self.chunked_prefill_size //= 4  # make it 2048
-            self.cuda_graph_max_bs = 4
 
         if not is_flashinfer_available():
             self.attention_backend = "triton"
@@ -194,7 +194,7 @@ class ServerArgs:
             self.cuda_graph_max_bs = min(self.cuda_graph_max_bs, 96)
             self.enable_overlap_schedule = False
             logger.warning(
-                f"DP attention is enabled. The chunked prefill size is adjusted to {self.chunked_prefill_size} to avoid MoE workload issue. "
+                f"DP attention is enabled. The chunked prefill size is adjusted to {self.chunked_prefill_size} to avoid MoE kernel issues. "
                 f"The CUDA graph max batch size is adjusted to {self.cuda_graph_max_bs}. "
                 "Data parallel size is adjusted to be the same as tensor parallel size."
             )
@@ -204,21 +204,8 @@ class ServerArgs:
                 "Overlap scheduler mode is enabled. This is an experimental feature. "
                 "Sampling penalizer (e.g., frequency and repetition penalty), constrained decoding (e.g., regex, JSON), "
                 "and embedding APIs are not supported and will lead to wrong results. "
-                "The NaN detection is also disabled."
             )
             self.disable_penalizer = True
-            self.disable_nan_detection = True
-
-        # Model-specific patches
-        if "Alibaba-NLP/gte-Qwen2-1.5B-instruct" == self.model_path:
-            logger.info(
-                "Not sure why, the tokenizer will add an additional token at the end of the prompt when trust_remote_mode=True"
-            )
-            self.trust_remote_code = False
-
-        if "gemma-2" in self.model_path.lower():
-            logger.info("When using sliding window in gemma-2, turn on flashinfer.")
-            self.attention_backend = "flashinfer"
 
     @staticmethod
     def add_cli_args(parser: argparse.ArgumentParser):
@@ -683,6 +670,11 @@ class ServerArgs:
             default=ServerArgs.torchao_config,
             help="Optimize the model with torchao. Experimental feature. Current choices are: int8dq, int8wo, int4wo-<group_size>, fp8wo",
         )
+        parser.add_argument(
+            "--enable-nan-detection",
+            action="store_true",
+            help="Enable the NaN detection for debugging purposes.",
+        )
         parser.add_argument(
             "--enable-p2p-check",
             action="store_true",